YouTube-VOS 2018 val, also known as YouTube Video Object Segmentation, is a video object segmentation dataset and the first test benchmark that supports multi-video task segmentation. The dataset contains 4,453 videos, of which 3,471 are used for training, 474 are used for validation, and 508 are used for testing. Every 5 frames of training and validation videos will have pixel-level ground truth annotations.

The dataset contains 7,800 independent objects, 190,000 high-quality manual annotations, and the video duration is more than 340 minutes.